Intelligent auditing data analyzing and processing method based on label self-evolution

Abstract

The invention discloses an audit data intelligent analysis processing method based on label self-evolution. Obtaining original data, processing the original data to obtain a structured feature set and a text vector set, performing standardization processing to obtain feature data, and storing the feature data in a feature bin; establishing an unsupervised calculation model, analyzing feature data by using the unsupervised calculation model, automatically identifying a potential risk mode, generating candidate tags, performing normalized weighted fusion to obtain a comprehensive score, adopting a reinforcement learning mechanism, presetting a reward function for the unsupervised calculation model and performing iterative updating, and introducing an incremental learning method to obtain a comprehensive score. Parameters of the unsupervised calculation model are updated in a small-step fast stack mode; a closed loop structure is formed by a data input layer, a label pool layer, an evolution engine layer and a feedback mechanism layer in an unsupervised calculation model. According to the method, automatic generation, quantitative evaluation, dynamic optimization and cross-scene migration can be realized, the manual maintenance cost is effectively reduced, and the adaptability and sustainable evolution capability of the model are improved.

Description

Technical Field

[0001] This invention specifically relates to an intelligent analysis and processing method for audit data based on tag self-evolution. Background Technology

[0002] In the field of intelligent auditing, the quality of data model construction directly determines the accuracy of risk identification and anomaly detection. Traditional intelligent auditing systems typically rely on manually defined tagging systems. Audit experts design tags based on policies, regulations, typical cases, and business experience to describe elements such as cash flow, contract execution, or project risks. These tags are then input into machine learning or rule engines to complete anomaly identification and risk alerts. However, this approach is highly dependent on expert experience and manual updates, making it difficult to adapt to the real-world needs of rapidly growing data volumes and frequent changes in the business environment. Because tagging systems usually exist in the form of static forms or dictionaries, their update cycles are long and their coverage is limited, making it difficult to capture the dynamic evolution of risk patterns in complex audit scenarios.

[0003] At the data processing level, traditional intelligent auditing models often employ fixed feature construction and indicator systems. Structured data is characterized by manually extracting fields and calculating indicators, while unstructured data such as text, contracts, and images rely on manual annotation or keyword matching to transform them into limited features. During model training, algorithms such as logistic regression, decision trees, and support vector machines are commonly used for classification or prediction, and their performance is highly dependent on the quality of feature selection and sample annotation. Due to the static definition of the labeling system and the lag in feature updates, these models often degrade rapidly after deployment, making it difficult to cope with data distribution drift, policy changes, or the emergence of new risk patterns. Furthermore, the labeling system faces significant migration barriers across auditing tasks in multiple scenarios. Different business lines use significantly different labeling standards, coding rules, and field definitions, resulting in a lack of a common semantic foundation for labels across different systems, preventing models from being reused across domains. Even within the same system, label versions from different time periods may have definition differences, further increasing data governance costs and model instability. To address this, existing systems often maintain label consistency through manual review or script alignment, but as the data scale increases, the manual reliance and maintenance costs of this method rise sharply.

[0004] A more prominent problem is that traditional labeling systems lack a quantitative evaluation mechanism for label effectiveness. Once labels enter the model input set, they are typically retained for a long time without performance verification or value updates, leading to the continuous accumulation of redundant labels and low-relevance features. This increases model complexity without a significant improvement in accuracy. Some studies have attempted to introduce label selection based on feature importance or statistical weights, but these still rely primarily on manually set thresholds, making it difficult to achieve adaptive selection. Furthermore, due to the lack of a continuous feedback mechanism, auditors' review conclusions, model false alarm samples, and system performance indicators fail to be fed back to the labeling management layer, preventing the model from self-correcting based on actual performance, resulting in a widespread signal lag problem.

[0005] In summary, traditional intelligent audit tagging systems rely primarily on static rules and human experience, with tag generation, evaluation, and optimization processes being fragmented. As multi-source heterogeneous data continues to grow, this model's shortcomings in tag timeliness, cross-scenario portability, and self-learning capabilities become increasingly apparent, making it difficult to support the high-performance audit analysis needs of complex environments and dynamic tasks. Summary of the Invention

[0006] Purpose of the invention: To provide an intelligent analysis and processing method for audit data based on tag self-evolution, which solves the above-mentioned problems existing in the prior art.

[0007] Technical solution: A method for intelligent analysis and processing of audit data based on tag self-evolution, comprising the following steps:

[0008] Obtain raw data, which includes at least general ledger, vouchers, accounts receivable and payable, budget execution, contract texts and invoice images. Process the raw data to obtain a structured feature set and a text vector set. Standardize the structured feature set and the text vector set respectively to obtain feature data and store it in the feature warehouse.

[0009] An unsupervised computation model is built to analyze feature data, automatically identify potential risk patterns, generate candidate labels, quantify the value of candidate labels based on feature importance and marginal contribution, and then perform normalized weighted fusion to obtain a comprehensive score. The candidate labels are then screened based on the comprehensive score.

[0010] A reinforcement learning mechanism is adopted to pre-set a reward function for the unsupervised computing model and iteratively update the label weights in the reward function. Incremental learning method is introduced to update the parameters of the unsupervised computing model in small steps and fast iterations.

[0011] The unsupervised computing model forms a closed-loop structure consisting of a data input layer, a label pool layer, an evolution engine layer, and a feedback mechanism layer. The system automatically collects the review conclusions of auditors, model operation indicators, and resource utilization information to form a feedback data stream. It then periodically re-optimizes the thresholds, weight coefficients, and model parameters to achieve continuous self-evolution of the label system.

[0012] Preferably, the raw data processing procedure is as follows:

[0013] First, the original data is uniformly encoded to obtain encoded data. After aligning the event records in the encoded data using a unified timestamp standard, the aligned encoded data is then detected and abnormal values ​​are removed based on the interquartile range method or the isolated forest algorithm. At the same time, at least sensitive fields such as name, account, and amount are desensitized using hash values, bucketing, or character replacement to obtain structured field data.

[0014] Preferably, the process for obtaining the structured feature set and text vector set is as follows:

[0015] Obtain structured field data, perform key feature filtering on the structured field data using Pearson correlation and mutual information to obtain filtered data, perform optical character recognition on the contract text and invoice images in the filtered data, and then generate text vectors from the remaining part of the filtered data using a BERT pre-trained model or a Doc2Vec pre-trained model to obtain a set of text vectors.

[0016] To obtain structured field data, at least aggregate statistics, divide intervals and perform ratio calculations on the amount, quantity, date and contract fields in the structured field data. At least construct a set of multi-dimensional numerical feature vectors that reflect the account balance and changes, transaction frequency, overdue days, budget execution deviation and contract execution progress indicators, and obtain a set of structured feature vectors.

[0017] Preferably, the candidate label includes at least attribute definition, semantic description, sample set, and source information, wherein the candidate label value quantification process is as follows:

[0018] The AUC (Area Under Curve) value is used to calculate the difference in performance between candidate labels and the unsupervised computation model. When the difference If the value is greater than 0, the label has a positive contribution to performance and is retained; if the difference is less than 0, the label is retained. When the value is approximately 0, the contribution is limited, so it is eliminated; when the difference is... If the value is less than 0, it is considered a noise label and will be removed, with a preset elimination threshold. When the difference <Elimination Threshold When the current candidate label is automatically eliminated, the difference is calculated. The calculation formula is as follows:

[0019] ;

[0020] In the formula: Indicates candidate tags, This represents an unsupervised computation model. This indicates that the tag has been removed. The model after that.

[0021] The preferred normalized weighted fusion process is as follows:

[0022] Feature importance scores and difference Unify the dimensions and normalize to obtain and Define the overall score:

[0023] ;

[0024] In the formula, Indicates control weight; This represents the normalized value of feature importance; a rolling window min-max normalization is used to smooth out extreme values, ensuring consistency between training and validation times during the evaluation process. and Perform a resampling stability test, and update the label status only when the direction is consistent and the confidence interval is stable, to ensure that the evaluation results are interpretable, reproducible and support the automated implementation of label selection.

[0025] Preferably, the calculation formula for the preset reward function is as follows:

[0026] ;

[0027] In the formula: This represents the improvement in the F1 score during the t-th iteration; F1 is a model performance metric, representing the harmonic mean of overall precision and recall. Indicates the change in false alarm rate. α represents computational and storage overhead; α, β, and γ are weighting coefficients.

[0028] Preferably, the label weight update process is as follows:

[0029] ;

[0030] In the formula: Indicates the first Wheel label weight, For learning rate; when Below the elimination threshold When that happens, remove the current candidate label.

[0031] Preferably, the process of updating the parameters of the unsupervised computation model through incremental learning and iterative small steps is as follows:

[0032] ;

[0033] In the formula: Indicates the first Wheel model parameters, For loss function, For new samples, the unsupervised computing model is updated using an incremental learning method, which allows it to absorb new data while maintaining existing knowledge, thereby reducing training overhead and improving adaptability.

[0034] Preferably, the data input layer is used to analyze feature data and automatically identify potential risk patterns; the label pool layer is used to store candidate labels and their attributes; the evolution engine layer integrates an unsupervised model, a reinforcement learning module, and an incremental learning module; and the feedback mechanism layer originates from the audit personnel's review results and the system operation logs, and is a feedback data stream automatically generated within the system.

[0035] Beneficial Effects: This invention relates to an intelligent analysis and processing method for audit data based on label self-evolution. By processing financial data to obtain a structured feature set and a text vector set, an unsupervised computation model is used to analyze the feature data, automatically identifying potential risk patterns and generating candidate labels. The value of the candidate labels is quantified based on feature importance and marginal contribution. The unsupervised computation model forms a closed-loop structure consisting of a data input layer, a label pool layer, an evolution engine layer, and a feedback mechanism layer. The system automatically collects auditor review conclusions, model operation indicators, and resource utilization information, forming a feedback data stream. Thresholds, weight coefficients, and model parameters are periodically re-optimized, achieving continuous self-evolution of the label system. This enables automatic generation, quantitative evaluation, dynamic optimization, and cross-scenario migration, effectively reducing manual maintenance costs and improving the model's adaptability and sustainable evolution capabilities. Attached Figure Description

[0036] Figure 1 This is a block diagram of the intelligent analysis and audit data system of the present invention. Detailed Implementation

[0037] like Figure 1 As shown, the present invention provides a technical solution: an intelligent analysis and processing method for audit data based on tag self-evolution, comprising the following steps:

[0038] The process involves acquiring raw data, including at least general ledger, vouchers, accounts receivable and payable, budget execution, contract texts, and invoice images. This raw data is then processed to obtain a structured feature set and a text vector set. The processing begins by uniformly encoding the raw data to obtain coded data. Event records in the coded data are then aligned using a unified timestamp standard. Anomalies are detected and removed from the aligned coded data based on the interquartile range (IQR) method or the isolated forest algorithm. Simultaneously, sensitive fields such as name, account number, and amount are desensitized using hash values, bucketing, or character replacement to obtain structured field data. The process for acquiring the structured feature set and text vector set is as follows:

[0039] Obtain structured field data, perform key feature filtering on the structured field data using Pearson correlation and mutual information to obtain filtered data, perform optical character recognition on the contract text and invoice images in the filtered data, and then generate text vectors from the remaining parts of the filtered data using a BERT pre-trained model or a Doc2Vec pre-trained model to obtain a text vector set. Obtain structured field data, and at least perform aggregation statistics, interval division and ratio calculation on the amount, quantity, date and contract fields in the structured field data to construct at least a multi-dimensional numerical feature vector set reflecting account balance and changes, transaction frequency, overdue days, budget execution deviation and contract execution progress indicators to obtain a structured feature vector set. After standardizing the structured feature set and the text vector set respectively, the feature data is obtained and stored in the feature warehouse.

[0040] An unsupervised computation model is constructed and used to analyze feature data, automatically identify potential risk patterns, and generate candidate labels. Each candidate label includes at least attribute definitions, semantic descriptions, sample sets, and source information. The process of quantifying the value of the candidate labels is as follows:

[0041] The AUC (Area Under Curve) value is used to calculate the difference in performance between candidate labels and the unsupervised computation model. When the difference If the value is greater than 0, the label has a positive contribution to performance and is retained; if the difference is less than 0, the label is retained. When the value is approximately 0, the contribution is limited, so it is eliminated; when the difference is... If the value is less than 0, it is considered a noise label and will be removed, with a preset elimination threshold. When the difference <Elimination Threshold When the current candidate label is automatically eliminated, the difference is calculated. The calculation formula is as follows:

[0042] ;

[0043] In the formula: Indicates candidate tags, This represents an unsupervised computation model. This indicates that the tag has been removed. The model then quantifies the value of candidate labels based on feature importance and marginal contribution, and performs normalized weighted fusion to obtain a comprehensive score. This comprehensive score is then used to screen candidate labels. The specific process of normalized weighted fusion is as follows:

[0044] Feature importance scores and difference Unify the dimensions and normalize to obtain and Define the overall score:

[0045] ;

[0046] In the formula, Indicates control weight; This represents the normalized value of feature importance; a rolling window min-max normalization is used to smooth out extreme values, ensuring consistency between training and validation times during the evaluation process. and Perform a resampling stability test, and update the label status only when the direction is consistent and the confidence interval is stable, to ensure that the evaluation results are interpretable, reproducible and support the automated implementation of label selection.

[0047] A reinforcement learning mechanism is employed to pre-define a reward function for the unsupervised computation model. The formula for calculating the pre-define reward function is as follows:

[0048] ;

[0049] In the formula: This represents the improvement in the F1 score during the t-th iteration; F1 is a model performance metric, representing the harmonic mean of overall precision and recall. Indicates the change in false alarm rate. This represents the computation and storage overhead; α, β, and γ are weight coefficients. After calculating the preset reward function, the label weights in the reward function are iteratively updated. The label weight update process is as follows:

[0050] ;

[0051] In the formula: Indicates the first Wheel label weight, For learning rate; when Below the elimination threshold At this time, the current candidate labels are removed, and incremental learning is introduced to update the parameters of the unsupervised computation model in small steps and iteratively. The process of updating the parameters of the unsupervised computation model in small steps and iteratively through incremental learning is as follows:

[0052] ;

[0053] In the formula: Indicates the first Wheel model parameters, For loss function, For new samples, the unsupervised computing model is updated using an incremental learning method, which allows it to absorb new data while maintaining existing knowledge, thereby reducing training overhead and improving adaptability.

[0054] The unsupervised computation model employs a closed-loop structure comprised of a data input layer, a label pool layer, an evolutionary engine layer, and a feedback mechanism layer. The data input layer analyzes feature data and automatically identifies potential risk patterns. The label pool layer stores candidate labels and their attributes, uniformly managing all label information and connecting the core hubs of generation, evaluation, and optimization / mitigation. During label generation, each label's attribute definition, semantic description, sample set, source information, generation batch, and algorithm fingerprint are written into the label pool, and a unique code is assigned to it. The label pool records the label's weight, comprehensive score, and current status (candidate, in use, downgraded, or eliminated), while maintaining version snapshots and lineage relationships. The system automatically generates change logs, enabling label traceability, rollback, and full lifecycle management. During operation, the label pool receives real-time weight adjustments, score updates, and elimination results from the evolutionary engine to ensure the consistency and manageability of label states. This layer provides read / write services to the evolution engine layer and feedback mechanism layer through standardized interfaces, ensuring data flow and state synchronization between different modules. The evolution engine layer integrates unsupervised models, reinforcement learning modules, and incremental learning modules. The feedback mechanism layer originates from auditor review results and system operation logs, and is an automatically generated feedback data stream within the system. The system automatically collects auditor review conclusions, model operation indicators, and resource utilization information to form a feedback data stream, which periodically re-optimizes thresholds, weight coefficients, and model parameters to achieve continuous self-evolution of the label system. The unsupervised computation model uses a sliding window algorithm to perform trend analysis on operation indicators, identify performance fluctuations and changes in label contributions, and automatically generate adjustment schemes for thresholds, learning rates, and reward coefficients. When a decline in model performance or data distribution drift is detected, the system triggers the evolution engine for local re-optimization through the feedback mechanism, or calls the version snapshot function of the label pool to perform a rollback to restore system stability. It consists of a data input layer, a tag pool layer, an evolution engine layer, and a feedback mechanism layer working together. Each layer interacts with information and synchronizes its state through a unified data interface, forming a closed-loop system of "input-generation-optimization-feedback" to achieve continuous learning and steady-state evolution of the system.

[0055] In this embodiment, data from the 2020-2022 audit project of the procurement of traditional Chinese medicine decoction pieces in a certain city hospital were collected. The traditional labeling system and the method proposed in this invention were applied to the project to conduct a comparative analysis of the procurement of a certain type of traditional Chinese medicine. The results obtained based on the problem characteristics of the collected project data and the model design logic are shown in Table 1 below.

[0056] Table 1. Performance Comparison of Traditional Tagging Systems and Tag Self-Evolution Frameworks

[0057]

[0058] The data anomaly identification rate of the self-evolving tag audit data analysis method based on the present invention can reach 84.3%, which can identify more abnormal procurement behaviors compared with the traditional tag system, and shows an improving trend in coverage and accuracy. The simulated improvement of AUC suggests that its discrimination ability in complex business data may be stronger. The simulated increase in the number of high-risk tags shows that the system has the ability to dynamically generate and retain more valuable tag features, reflecting its practical application value.

[0059] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for intelligent analysis and processing of audit data based on tag self-evolution, characterized in that, Includes the following steps: Obtain raw data, which includes at least general ledger, vouchers, accounts receivable and payable, budget execution, contract texts and invoice images. Process the raw data to obtain a structured feature set and a text vector set. Standardize the structured feature set and the text vector set respectively to obtain feature data and store it in the feature warehouse. An unsupervised computation model is built to analyze feature data, automatically identify potential risk patterns, and generate candidate labels. The value of the candidate labels is quantified based on feature importance and marginal contribution. After normalization and weighted fusion, a comprehensive score is obtained, and the candidate labels are screened based on the comprehensive score. A reinforcement learning mechanism is adopted to pre-set a reward function for the unsupervised computing model and iteratively update the label weights in the reward function. An incremental learning method is introduced to update the parameters of the unsupervised computing model in small steps and fast iterations. The unsupervised computing model forms a closed-loop structure consisting of a data input layer, a label pool layer, an evolution engine layer, and a feedback mechanism layer. The system automatically collects the review conclusions of auditors, model operation indicators, and resource utilization information to form a feedback data stream. It then periodically re-optimizes the thresholds, weight coefficients, and model parameters to achieve continuous self-evolution of the label system.

2. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 1, characterized in that, The processing procedure for the raw data is as follows: First, the original data is uniformly encoded to obtain encoded data. After aligning the event records in the encoded data using a unified timestamp standard, the aligned encoded data is then detected and abnormal values ​​are removed based on the interquartile range method or the isolated forest algorithm. At the same time, at least sensitive fields such as name, account, and amount are desensitized using hash values, bucketing, or character replacement to obtain structured field data.

3. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 2, characterized in that, The process of obtaining the structured feature set and text vector set is as follows: Obtain structured field data, perform key feature filtering on the structured field data using Pearson correlation and mutual information to obtain filtered data, perform optical character recognition on the contract text and invoice images in the filtered data, and then generate text vectors from the remaining part of the filtered data using a BERT pre-trained model or a Doc2Vec pre-trained model to obtain a set of text vectors. To obtain structured field data, at least aggregate, divide, and perform ratio calculations on the amount, quantity, date, and contract fields in the structured field data. At least construct a set of multi-dimensional numerical feature vectors that reflect account balance and changes, transaction frequency, overdue days, budget execution deviation, and contract execution progress indicators to obtain a set of structured feature vectors.

4. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 1, characterized in that, The candidate labels include at least attribute definitions, semantic descriptions, sample sets, and source information, wherein the candidate label value quantification process is as follows: The AUC (Area Under Curve) value is used to calculate the difference in performance between candidate labels and the unsupervised computation model. When the difference If the value is greater than 0, the label has a positive contribution to performance and is retained; if the difference is less than 0, the label is retained. When the value is approximately 0, the contribution is limited, so it is eliminated; when the difference is... If the value is less than 0, it is considered a noise label and will be removed, with a preset elimination threshold. When the difference <Elimination Threshold When the current candidate label is automatically eliminated, the difference is calculated. The calculation formula is as follows: ； In the formula: Indicates candidate tags, This represents an unsupervised computation model. This indicates that the tag has been removed. The model after that.

5. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 3, characterized in that, The specific process of normalized weighted fusion is as follows: Feature importance scores and difference Unify the dimensions and normalize to obtain and Define the overall score: ； In the formula, Indicates control weight; This represents the normalized value of feature importance; a rolling window min-max normalization is used to smooth out extreme values, ensuring consistency between training and validation times during the evaluation process. and Perform a resampling stability test, and update the label status only when the direction is consistent and the confidence interval is stable, to ensure that the evaluation results are interpretable, reproducible and support the automated implementation of label selection.

6. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 1, characterized in that, The formula for calculating the preset reward function is as follows: ； In the formula: This represents the improvement in the F1 score during the t-th iteration; F1 is a model performance metric, representing the harmonic mean of overall precision and recall; Indicates the change in false alarm rate. α represents computational and storage overhead; α, β, and γ are weighting coefficients.

7. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 2, characterized in that, The tag weight update process is as follows: ； In the formula: Indicates the first Wheel label weight, For learning rate; when Below the elimination threshold When that happens, remove the current candidate label.

8. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 1, characterized in that, The process of updating the parameters of an unsupervised computation model through incremental learning and iterative small steps is as follows: ； In the formula: Indicates the first Wheel model parameters, For loss function, For new samples, the unsupervised computing model is updated using an incremental learning method, which allows it to absorb new data while maintaining existing knowledge, thereby reducing training overhead and improving adaptability.

9. The intelligent analysis and processing method for audit data based on tag self-evolution as described in claim 1, characterized in that, The data input layer is used to analyze feature data and automatically identify potential risk patterns. The label pool layer is used to store candidate labels and their attributes. The evolution engine layer integrates unsupervised models, reinforcement learning modules, and incremental learning modules. The feedback mechanism layer is derived from the audit personnel's review results and system operation logs, and is a feedback data stream automatically generated within the system.

Images